Towards Scalable Image Classifier Learning with Noisy Labels via Domain Adaptation

Abstract

In this chapter we focus on learning image classifiers with noisy labels through domain adaptation. Existing approaches for learning image classifiers with noisy labels using human supervision are generally difficult to scale to large set of classes as manual labeling images for all classes are expensive and time-consuming. Approaches that address noisy labels without manual labeling efforts are scalable but less effective in lack of reliable supervision. Transfer learning reconciles this conflict through transferring knowledge from classes with exemplary human supervision (source domains) to classes where data are not manually verified (target domains), relaxing the requirement of human efforts. In this chapter, we introduce a transfer learning set-up for tackling noisy labels, and review CleanNet, the first neural network model that practically implements this set-up, and explore future directions of this topic.